Energy Consumption Minimization in MEC-Empowered CF-mMIMO System Based on Multi-Agent Deep Reinforcement Learning

To satisfy strict latency requirements, we consider a mobile edge computing (MEC)-empowered cell-free massive multiple-input multiple-output (CF-mMIMO) system. To minimize the total energy consumption of user ends under low-latency constraints, an optimization problem is formulated that combines the optimization of offloading decisions, power allocation, and computing resource allocation. By considering the distributed architecture of CF-mMIMO, a partially centralized training multi-agent proximal policy optimization (PC-MAPPO) algorithm is proposed. This algorithm enables the coexistence of both centralized and distributed training in the system, reducing computational complexity while saving communications overhead. Meanwhile, the scheme allows for the existence of different types of agents to jointly optimize the actions of users and access points (APs). The simulation results show that the proposed algorithm reduces energy consumption compared to the considered traditional solutions, while having lower computational complexity and better scalability.